NVIDIA lists October 28, 2026, as the release date for Nemotron 3.5 Transcribe. Its model card describes a 34B-parameter offline speech model that handles recordings up to one hour, with supported tasks varying by language.
What Nemotron 3.5 Transcribe does
Nemotron 3.5 Transcribe is a speech-augmented language model, or SALM. NVIDIA says it can produce verbatim or cleaned-up transcripts, identify speakers, add timestamps, translate speech, identify languages and answer questions about audio. The full model has 34 billion parameters, including about 4.3 billion active parameters; its language backbone is Nemotron 3.5 Lightning 30B-A3B.
Language coverage varies by task
NVIDIA lists 40 languages, but not every language supports every task. Its model card, for example, lists Tamil and Turkish for transcription but not translation. Audio understanding and reasoning were evaluated only in English.
Offline operation and deployment
This model processes recordings offline, not as a real-time streaming service, and NVIDIA specifies a maximum recording length of one hour. A separate 0.6B Nemotron 3.5 ASR model is the streaming-oriented option.
NVIDIA’s reference deployment uses one H100 80GB GPU. Its compatibility table lists other hardware and precision combinations, including the RTX 5090 with NVFP4, so the reference setup is not a universal hardware requirement. NVIDIA identifies OpenMDW-1.1 as the governing license and permits commercial and non-commercial use under its terms. The NVIDIA model card gives the deployment and license details.
